AICTE Pragati Scholarship Assistant

A bilingual, source-grounded QA system for AICTE Pragati Scholarship guidelines. Answers in English or Hindi using official docs with verified source citations.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for AICTE Pragati Scholarship A## Inspiration

Eligibility rules, amounts, and deadlines for scholarships like the AICTE Pragati Scholarship are often buried in long, complex PDFs primarily available in English. Students and families more comfortable in Hindi frequently struggle to find simple answers—such as family income limits or yearly grant amounts.

While general AI chatbots can answer from general knowledge, relying on hallucinations in a financial and eligibility context can mislead students. This inspired us to build an assistant governed by three strict rules:

  1. Answer only from the official document.
  2. Show the source for every answer with page numbers.
  3. Refuse safely instead of guessing when the document lacks the answer.

What it does

  • Bilingual Support: Accepts questions and responds accurately in both English and Hindi.
  • Source-Grounded QA: Retrieves relevant sections of the guideline PDF and provides direct citations with page numbers so users can verify.
  • Safe Refusals: Politely responds with "I could not find this in the guideline, so I won't guess" if the answer is absent from the official source.
  • Transparency: Always highlights the document date, ensuring users are aware if guidelines change.

How we built it

  1. Extraction: Extracted text from the official guideline PDF using PyMuPDF while preserving page numbers.
  2. Chunking Strategies: Tested two chunking methods:
    • Section chunks: One chunk per numbered section plus annexures (12 chunks total).
    • Fixed chunks: 120-word windows with a 30-word overlap (15 chunks total).
  3. Retrieval Pipeline: Utilized a robust hybrid search combining:
    • BM25 for keyword matching.
    • Dense embeddings using multilingual-e5-small for cross-lingual (Hindi/English) semantic understanding.
    • Reciprocal Rank Fusion (RRF) to merge rankings.
  4. Generation: Powered by Gemini (gemini-3.5-flash-lite), passing top-retrieved sections with strict instructions to restrict context, maintain exact numerical data, and reply in the user's language.
  5. Interface: Built an interactive web demo using Streamlit.

Evaluation & Findings

We evaluated the system on a test set of 33 questions (25 English, 4 Hindi, 4 unanswerable edge cases).

  • Section Chunks Outperform Fixed Chunks: Splitting documents at natural section boundaries keeps related facts together, boosting retrieval accuracy significantly (Dense hit@3: 0.96 vs 0.84 for fixed chunks).
  • Hybrid Retrieval Wins: Hindi queries heavily benefited from dense embeddings (keyword-only search struggled with exact matching on Hindi, scoring 0.25, whereas Hybrid + Section chunks achieved a stellar 1.00 hit@3).
  • Reliable Refusals: Successfully handled out-of-scope queries (like current-year application deadlines absent from older documents) by defaulting to safe refusals.

Challenges we ran into

  • Cross-Lingual Limitations: Traditional keyword search failed completely on Hindi queries unless exact English terms were matched, which drove our implementation of a multilingual dense retrieval model.
  • Context Splitting: Fixed-size chunking frequently split critical tables (like state-wise scholarship breakdowns) across chunks, leading us to adopt logical section-based chunking.
  • API Quota Management: Managing Gemini free-tier rate limits during extensive iterative testing required optimizing our pipeline calls.

What we learned

We learned the critical importance of deterministic grounding in high-stakes domains like scholarship applications. Optimizing hybrid search tuning and handling multilingual embeddings effectively bridged the gap between raw document complexity and end-user accessibility.

What's next for AICTE Pragati Scholarship Assistant

  • Multi-Scheme Expansion: Incorporate additional schemes (Pragati Diploma, Saksham) with automated version tracking for updated guidelines.
  • Interactive Eligibility Checker: Build a quick questionnaire form (income, course year, state) to instantly map students to schemes they qualify for.
  • Crowdsourced Test Expansion: Grow the evaluation dataset with real student queries, incorporating mixed Hinglish prompts and human evaluation loops.

Built With

Share this project:

Updates

Submission history